Harnessing Machine Learning and Physiological Knowledge for a Novel EMG-Based Neural-Machine Interface.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Neuroscience

Background:

• Decoding continuous joint angles from electromyography (EMG) signals is crucial for advanced prosthetic control.
• Existing methods, such as musculoskeletal models (MM) and artificial neural networks (ANNs), have limitations in robustness and accuracy.
• A hybrid approach integrating machine learning and biomechanical principles may overcome these limitations.

Purpose of the Study:

• To develop and evaluate a novel hybrid electromyography-based neural machine interface (NMI) called the Neural Network-Musculoskeletal hybrid Model (N2M2).
• To decode continuous joint angles by combining machine learning and musculoskeletal modeling concepts.
• To assess the performance of N2M2 against traditional MM and ANNs (MLP, NARX) under various conditions.

Main Methods:

• Collected EMG and joint kinematics data from 10 non-disabled and 1 transradial amputee subject.
• Developed the N2M2 by integrating machine learning with musculoskeletal modeling.
• Compared N2M2 performance against MLP, NARX, and MM through offline (posture variations, electrode shifts, noise) and online (postural matching task) evaluations.

Main Results:

• N2M2 demonstrated superior prediction accuracy compared to MLP across different postures and electrode locations (p < 0.003).
• N2M2 showed reduced sensitivity to noisy EMG signals for joint angle estimation versus MM and NARX (p < 0.032 for error, p = 0.007 for correlation).
• N2M2 achieved better online task performance than the NARX network (p ≤ 0.030).

Conclusions:

• Combining machine learning and musculoskeletal modeling yields a more robust joint kinematics decoder than either approach alone.
• The N2M2 represents a significant advancement in EMG-based neural machine interfaces.
• This hybrid model holds promise for developing highly reliable controllers for powered prosthetic hands.